AI agents hallucinate package names. need gives them a verified index of 10,000+ CLI tools — and a closed feedback loop that gets smarter with every install.

What happens

You ask Claude to "compress these PNGs". Claude doesn't have pngquant installed and doesn't know what the best tool is. But need is running as an MCP server in the background, so Claude automatically:

1. Searches need for "compress png images"
2. Installs the top result (brew install pngquant)
3. Runs it on your files
4. Reports that it worked — so the next agent's search ranks pngquant higher

You never interact with need directly. You just see the result.

  search → install → use → report
    ↑                        |
    └────── rankings ────────┘

Install

npm install -g @agentneeds/need

That's it. MCP servers are automatically configured for Claude Code, Cursor, and Claude Desktop on install. Your AI agent can immediately discover and install CLI tools without you doing anything.

Or run with npx: npx @agentneeds/need "compress png images"

How agents use it

Under the hood, need exposes three MCP tools that agents call autonomously:

1. search_tools — semantic search across 10,000+ CLI tools
2. install_tool — install the best match (security allowlist: brew, apt, npm, pip, cargo only)
3. report_tool_usage — report success or failure, improving rankings for every future agent

No API keys. No accounts. No configuration.

Setup

MCP is configured automatically on install. To manually reconfigure:

need setup

To add project-level config files for GitHub Copilot and Windsurf:

need init

Works for humans too

need also works as a standalone CLI — semantic search that understands intent, not just keywords.

need convert pdf to png
need find duplicate files
need compress video without losing quality

How it works

Queries are embedded and matched against a pgvector database of CLI tools. Results are ranked by semantic similarity combined with community success/failure signals from report_tool_usage.

Links

• GitHub
• Website
• npm

License

MIT